Three Hundred Years of Monthly Return Predictability : A Comprehensive Examination and Transfer Learning

I construct a 300-year monthly testing sample for the UK market and study the return predictability in both the UK and the US. I conduct out-of-sample tests with 312 prediction setups based on 23 popular predictors, 3 model updating windows, and 3 common forecast combinations. Over the long run, the predictability declines substantially. Only 12 setups show out-of-sample R-Squared greater than 0.01 in the UK, while 19 setups have out-of-sample R-Squared greater than 0.01 in the US. 23 setups lead to positive but limited economic gains in the UK. However, most setups realize sizable economic gains in the US. Transfer learning setups that fit models using UK data show substantially stronger performance in the US market, compared to fitting models using US data. The predictability shows tremendous fluctuation in the short term, which urges caution in interpreting out-of-sample tests. The short-term fluctuation of predictability can easily lead to disagreements in testing results due to the change of sample coverage. With the UK history, I show that the predictability concentrates in interaction periods between GDP turning points and extreme return months. The predictability also concentrates in rare events, such as epidemic years. On average, the predictability is the weakest in summer and the strongest in winter and spring